Skip to main content

Ant Group Open-Sources LingBot-Vision: Giving Robots a Sense of Space

In the quest to make robots perceive the physical world as accurately as humans, a new breakthrough has emerged from Ant Group's embodied intelligence arm, Robbyant. They've just open-sourced the LingBot-Vision model family, a set of self-supervised vision Transformers that take a fresh approach to spatial understanding.

Most visual AI models today are obsessed with object recognition—they're great at telling you what's in an image, but they often miss the finer details crucial for physical interaction: boundaries, contours, and depth. LingBot-Vision flips this priority on its head. Instead of focusing on what objects are, it zeroes in on where they begin and end. The key innovation is "mask boundary modeling," a technique that trains the model to identify the most information-rich boundary regions in an image and use them as the core learning signal. This allows the model to not only understand semantics but also develop a strong geometric sense of space.

Image

And the results speak for themselves. The flagship model, ViT-g/16, packs only 1.1 billion parameters, yet it achieved top performance on the NYU-Depth v2 depth estimation benchmark. It even surpassed DINOv3, a model with 7 billion parameters, while using a training dataset only about one-third the size. For those with limited computing resources, the series also offers distilled versions ranging from 300 million parameters down to smaller sizes, ensuring leading dense prediction performance across different hardware.

To showcase the real-world value, the team also upgraded their depth completion system, LingBot-Depth 2.0. Testing shows it significantly improves accuracy when handling transparent objects—a traditional blind spot for perception systems. As data scales up, LingBot-Vision's performance continues to improve without the saturation seen in traditional models, highlighting the potential of this boundary-centric approach for complex environments.

Currently, LingBot-Vision is fully open-sourced on Hugging Face under the Apache-2.0 license, including weights and inference code for four model sizes. With this technology, developers can give robots more sensitive physical perception at lower computational costs, pushing embodied intelligence toward a more precise interactive future.

Key Points

  • Innovative Approach: LingBot-Vision uses mask boundary modeling to prioritize object boundaries and contours over traditional object recognition.
  • Impressive Performance: The 1.1B parameter ViT-g/16 model outperforms larger models like DINOv3 (7B params) on depth estimation tasks.
  • Practical Deployment: Distilled versions from 300M parameters enable deployment on resource-constrained hardware.
  • Real-World Application: The upgraded LingBot-Depth 2.0 system improves accuracy for transparent objects, a common challenge.
  • Open Source: Fully available on Hugging Face under Apache-2.0 license, including weights and inference code.